2002, Object Detection Using the Statistics of Parts

@article{schneiderman2004odu, author = {Schneiderman, H. and Kanade, T. }, citeulike-article-id = {1536372}, comment = {A journal paper summarizing the work of Schneiderman and Kanade on object detector. Their approach is a little bit outperformed by Viola-Jones cascade, which is in general much faster. Jet, this paper offers some interesting directions, which might be used even in WaldBoost. The detector is based on statistics of 'parts'. Each part is a subset of wavelet coefficients, these coefficients have high mutual information. The wavelet response is discretized and the joint distribution is build using table. Several 'parts' are selected during training, which uses Real AdaBoost. Corbis is an example of online picture database with growing number of entries : there is a obvious need for effective object retrieval. The detector factors out the pose variability - it trains in fact two detecrs (frontal, profile) for faces and eight detecectors for cars. ??? Do we realy need to simplify the task so much? It would be much convenient to have one encapsulating algorithm. The detection is performed by scanning the classifier over the original image and a series of resized versions. Naive approach to detection would be to model the 20x20x255x2 distribution by table, which is in fact not feasible because of lack of examples -> some constraint has to be applies. There are two basic approaches: global and part-based. Global approach can take only limited family of functions (linear - SVM, LDA, quadratic), or reduce the dimensionality of data (PCA). Part-based approach model some group of pixels or transformed variables (wavelet coefficients) more accurately (e.g. tables) but these parts are considered independent from each other -> simple combination of probability. Variables can be reused in multiple parts. Assumption: for a given object, each pixel is statistically related with some pixels more than others -> global model does not make good use of modeling resources (all dimensions have the same weight even if useful or invariant). ??? SVM on a face window with large background wouldn't work well? Part-based approach allocates richer models to stronger relationships, which we model by absolutely general tables. ??? In Adaboost - these strong relationships do not occur, because in every stage the best feature (Haar wavelet) is selected and in next step the most independent one. ??? How does it look the mutual information of features selected in AdaBoost? Wavelet basis are used 'decorrelation' (whitening the signal) - also called sparse representation. An approach similar to perception of cortical neurons. 5/3 linear phase wavelet filterbank. ??? Mutual information could be used for measuring similarity of Haar-features. We would search only those features which bring some additional inforamtion to classification. Wavelet coeffitients are able to represent small regions over high fequencies, larger regions over low frequencies, horizontal, vertical inforamtion -> wavelet transformation allowe this. !!! Information about parts with low energy (low variance) are discriminative. If we have 'representative' data, we could simply model P(part|object). This is overcome by training the classifier to minimize classification error on training set -> AdaBoost. ??? How to use already computed Haar-features for neighbouring window? }, journal = {International Journal of Computer Vision}, keywords = {2002, 500, car, cmu, detection, face, ijcv, object, wavelets}, number = {3}, pages = {151--177}, posted-at = {2007-08-05 13:15:53}, priority = {5}, publisher = {Springer}, title = {2002, Object Detection Using the Statistics of Parts}, volume = {56}, year = {2004} }

TY - JOUR ID - schneiderman2004odu L3 - citeulike-article-id:1536372 TI - 2002, Object Detection Using the Statistics of Parts JF - International Journal of Computer Vision VL - 56 IS - 3 SP - 151 EP - 177 PB - Springer KW - 2002 KW - 500 KW - car KW - cmu KW - detection KW - face KW - ijcv KW - object KW - wavelets N1 - A journal paper summarizing the work of Schneiderman and Kanade on object detector. Their approach is a little bit outperformed by Viola-Jones cascade, which is in general much faster. Jet, this paper offers some interesting directions, which might be used even in WaldBoost. The detector is based on statistics of 'parts'. Each part is a subset of wavelet coefficients, these coefficients have high mutual information. The wavelet response is discretized and the joint distribution is build using table. Several 'parts' are selected during training, which uses Real AdaBoost. Corbis is an example of online picture database with growing number of entries : there is a obvious need for effective object retrieval. The detector factors out the pose variability - it trains in fact two detecrs (frontal, profile) for faces and eight detecectors for cars. ??? Do we realy need to simplify the task so much? It would be much convenient to have one encapsulating algorithm. The detection is performed by scanning the classifier over the original image and a series of resized versions. Naive approach to detection would be to model the 20x20x255x2 distribution by table, which is in fact not feasible because of lack of examples -> some constraint has to be applies. There are two basic approaches: global and part-based. Global approach can take only limited family of functions (linear - SVM, LDA, quadratic), or reduce the dimensionality of data (PCA). Part-based approach model some group of pixels or transformed variables (wavelet coefficients) more accurately (e.g. tables) but these parts are considered independent from each other -> simple combination of probability. Variables can be reused in multiple parts. Assumption: for a given object, each pixel is statistically related with some pixels more than others -> global model does not make good use of modeling resources (all dimensions have the same weight even if useful or invariant). ??? SVM on a face window with large background wouldn't work well? Part-based approach allocates richer models to stronger relationships, which we model by absolutely general tables. ??? In Adaboost - these strong relationships do not occur, because in every stage the best feature (Haar wavelet) is selected and in next step the most independent one. ??? How does it look the mutual information of features selected in AdaBoost? Wavelet basis are used 'decorrelation' (whitening the signal) - also called sparse representation. An approach similar to perception of cortical neurons. 5/3 linear phase wavelet filterbank. ??? Mutual information could be used for measuring similarity of Haar-features. We would search only those features which bring some additional inforamtion to classification. Wavelet coeffitients are able to represent small regions over high fequencies, larger regions over low frequencies, horizontal, vertical inforamtion -> wavelet transformation allowe this. !!! Information about parts with low energy (low variance) are discriminative. If we have 'representative' data, we could simply model P(part|object). This is overcome by training the classifier to minimize classification error on training set -> AdaBoost. ??? How to use already computed Haar-features for neighbouring window? AU - Schneiderman, H AU - Kanade, T PY - 2004/// ER -